AdaDurIAN: Few-shot Adaptation for Neural Text-to-Speech with DurIAN

TL;DR

AdaDurIAN introduces a novel few-shot learning approach for neural TTS that leverages a DurIAN-based model to improve pronunciation accuracy and cross-lingual fluency with limited data, outperforming end-to-end systems.

Contribution

The paper proposes AdaDurIAN, a DurIAN-based model that enables effective few-shot speaker adaptation and emotion transfer in neural TTS, addressing alignment issues in end-to-end models.

Findings

AdaDurIAN outperforms baseline end-to-end systems in few-shot TTS tasks.

Subjective evaluations show higher naturalness and speaker similarity scores.

Demonstrates promising results in emotion transfer applications.

Abstract

This paper investigates how to leverage a DurIAN-based average model to enable a new speaker to have both accurate pronunciation and fluent cross-lingual speaking with very limited monolingual data. A weakness of the recently proposed end-to-end text-to-speech (TTS) systems is that robust alignment is hard to achieve, which hinders it to scale well with very limited data. To cope with this issue, we introduce AdaDurIAN by training an improved DurIAN-based average model and leverage it to few-shot learning with the shared speaker-independent content encoder across different speakers. Several few-shot learning tasks in our experiments show AdaDurIAN can outperform the baseline end-to-end system by a large margin. Subjective evaluations also show that AdaDurIAN yields higher mean opinion score (MOS) of naturalness and more preferences of speaker similarity. In addition, we also apply AdaDurIAN to emotion transfer tasks and demonstrate its promising performance.